CHAPTER 8
CONCLUSIONS
8.1 General Conclusions 7 a
In this book, we have attempted to give, at a fairly advanced level
tor "a unifed treatment of current methodologies for the design and
analysis of adaptive control algorithms, ;
First, we presented several schemes for the adaptive identification
and contol of linear time invariant systems, An output eror scheme,
an input error scheme, and an indirect scheme were derived in a unifi
framework Whi al the schemes were shown tobe globally sable, te
assumptions that went into the derivation of the schemes were quite
different. For instance, the input error adaptive cortral scheme did i
require a strictly positive real (SPR) condition for the reference mode
‘This also had implications for the transient behavior of the adaptive sys-
tems. ;
‘A malo gol of this book bas been the presentation of a number of
recent techniques for analyzing the stability, parameter convergence ar
robusines of the complinied nonlinear dynamics herent i the adap,
tive algorithms. For the stability proofs, we presented a sequence of
lemmas drawn from the literature on input-output L, stability. For the
parameter convergence proofs, we used results from generalized har-
analysis, and extracted frequency-domain conditions. For the
study of robustness, we exploited Lyapunov and averaging methods, We
feel that a complete mastery of these techniques will lay the groundwor
for future studies of adaptive systems.
324
Section 8.1 General Conclusions 328
While we did not deal explicitly with discrete time systems, our
Bresentation of the continuous time results may be transcribed to the
diserete time case with not much dificulty. The operator relationships
that were used for continuous time systems (L, spaces) also hold true for
discrete time systems (l, spaces). In fact, many derivations may be
simplified in the discrete time case because continuity conditions (euch
4 the regularity of signals) are then automaticaly satished
‘Averaging techniques have proved extremely useful and it is likely
that important developments will stil follow from their use. It iy
interesting to note that the two-time scale approximation was not only
fundamental to the application of averaging methods to convergence
(Chapter 4) and to robustness (Chapter 5), but was also underlying in the
Proofs of ‘exponential convergence (Chapter 2), and plobal. stability
(Chapter 3). This highlights the separation between adaptation and com,
trol, and makes the connections between direct and indirect adaptive
control more obvious.
Methods for the analysis of adaptive systems were a focal point of
this book. As was observed in Chapter 5, algorithms that are stable for
some inputs may be unstable for others, While simulations, are
{fatremely valuable to illustrate a point, they are useless to prove any glo,
bal behavior of the adaptive algorithm. This is a crucial consequence of
the nonlinearity of the adaptive systems, that makes rigorous analysis
techniques essential to progress inthe area,
82 Future Research
Adaptive control is a very active area of research, and there is a great
deal more to be done. The area of robustness is essential to successful
applications, and since the work of Rohrs et al, it has been understood
that the questions of robustness for adaptive systems are very different
fiom the same questions for linear time-invariant systems. This is due
in great part to the dual control aspect of adaptive systems: the refer.
face input plays a role in determining the convergence and robustness
by providing excitation to the identification loop. A major problem
Fomains {o quantify robustness for adaptive systems, Current theory
d2¢s not allow for the comparison of the robustness of different adaptive
systems, and the relation to non-adaptive robustness concepts. Closer
fonnections will probably emerge from the application of averaging
methods, and from the-frequency-domain results that they lead to.
Besides these fundamental questions of analysis, much remains to
te done to precisely define design methodologies for robust adaptive sys.
lems and in particular a better understanding of which algorithms sec
more robust. Indeed, although the adaptive systems discussed in this
book have identical stability properties in the ideal case, there ic326 Conclusions Chapter 8
evidence that their behavior is drastically different in the presence of
unmodeled dynamics. A better understanding of which algorithms are
‘more robust will also help in deriving guidelines for the improved design
of robust algorithms,
While we have extensively discussed the analysis of adaptive sys-
tems, we also feel that great strides in this area will come from experi
ences in implementing the algorithms on several classes of systems.
With the advent of microprocessors, and of today’s multi-processor
environments, complicated algorithms can now be implemented at very
high sample rates. The years to come will see a proliferation of tech-
niques to effectively map these adaptive algorithms onto multiprocessor
control architectures. There is a great deal of excitement in the control
community at large over the emergence of such custom multiprocessor
control architectures as CONDOR (Narasimhan et al [1988]) and
NYMPH (Chen er af [1986]). In turn, such advances will make it possi-
ble to exploit adaptive techniques on high bandwidth systems such as
exible space structures, aircraft flight control systems, light weight robot
manipulators, and the like. While past successes of adaptive control
have been on systems of rather low bandwidth and benign dynamics, the
Future years are going to be ones of experimentation on more challenging
systems.
‘Two other areas that promise explosive growth in the years to come
are adaptive control of multi-input multi-output (MIMO) systems, and
adaptive control of nonlinear systems, explicitly those linearizable by
state feedback, We presented in this book what we feel is the tip of the
iceberg in these areas. More needs to be understocd about the sort of
prior information needed for MIMO adaptive systems. Conversely, the
incorporation of various forms of prior knowledge into black-box models
of MIMO systems also needs to be studied. Adaptive control for MIMO
systems is especially attractive because the traditional and heuristic tech-
niques for SISO systems quickly fall apart when strong cross-couplings
appear. Note also that research in the identification of MIMO systems
is also relevant to nonadaptive algorithms, which are largely dependent
on the knowledge of a process model, and of its uncertainty. One may
hope that the recently introduced averaging techniques will help to
better connect the frequency-domain properties of adaptive and nona-
daptive systems,
A very large class of nonlinear systems is explicitly linearizable by
state feedback. The chief difficulty with implementing the linearizing
control law is the imprecise knowledge of the nonlinear functions in the
dynamics, some of which are often specified in table look-up form.
‘Adaptation then has a role in helping identify the nonlinear functions
on-line to obtain asymptotically the correct linearizirg control law. This
Section 8.2 Future Research a
Approach was discussed in this book, bu itis stl in its early develop-
ment. However, we have found it valuable inthe implementtioe alg
siapive contiler fran india root Cie "Adepel) ant
curently working on 2 Bight control sistem for 2 vena! heat
landing aircraft (the Harrier), - Solr and
In adtion to al these exciting new directions of research in adap-
tive control, most of which ae logal extensions and outgiowite ofthe
developments presented inthe previous chapters, we som pve fee
ather new vistas which are not ay abvious extensions,
‘A “Universal” Theory of Adaptive Control
While all the adaptive control algorithms developed in this book
required assumptions on the plant—in the single-input single-output
case, the order of the plant, the relative degree of the plant, the sign of
the high-frequency gain, and the minimum phase property of the plant
itis interesting to ask if these assumptions are a minimal set of assump.
tions. Indeed, that these assumptions can be relaxed was established by
Morse (1985, 1987], Mudgett and Morse [1985], Nussbaum {1983}, and
Martenson [1985] among others. Chief under the assumptions that
could be relaxed was the one on the sign of the high-frequency gain,
There is a simple instance of these results which is in some sense
Fepresentative of the whole family: consider the problem of adaptively
Labitzing a fst order linear plant of relative degree with unknown
in ky i,
Jp = Ay Yp + kyu 6.2.1)
With y different from zero but otherwise unknown, and dy unknown, If
the sign of k, is known and assumed positive, the adaptive control law
1 dly (8.2.2)
and
dy = yp (8.2.3)
gan be shown to yield yp 0 as 1 +00, Nussbaum [1983] proposed
that ifthe sign of ky is unknown, the control law (8.2.2) can be replaced
y
4 = dB (0) 005 (de(t))»p (8.2.4)
with (8.2.3) as before, He then showed that yp» 0 as ¢ + co, with
dd) remaining bounded. Heuristically, the feedback gain d3
ristically, the feedback gain d? cos(do) of
(82.4) alternates in sign (“searches for the correct sign”) as dp is328 Conclusions Chapter 8
decreased monotonically (by (8.2.3)) until it is large enough and of the
"correct sign” to stabilize the equation (8.2.1).
While the transient behavior of the algorithm (8.2.3), (8.2.4) is
poor, the scheme has stimulated a great deal of interest to derive adap-
tive control schemes requiring a minimal set of assumptions on the plant
(universal controllers). A further objective is to develop a unified frame-
work which would subsume all the algorithms presented thus far. Adap-
tive systems may be seen as the interconnection of a plant, a parameter-
ized controller, and adaptation law or tuner (cf. Morse [1988)). The
parameterized controller is assumed to control the process, and the tuner
assumed to tune the controller. Tuning is said to have taken place when
a suitable tuning error goes to zero. The goal of a universal theory is 10
sive @ minimal set of assumptions on the process, the parameterized
controller, and the tuner to guarantee global stability and asymptotic
performance of the closed loop system, Further, the assumptions are to
contain as special cases the algorithms presented thus far. Such a theory
would be extremely valuable from a conceptual and intellectual stand-
point.
Rule-Based, Expert and Learning Control Systems
As the discussions in Chapter 5 indicated, there is a great deal of work
needed to implement a given adaptive algorithm, involving the use of
heuristics, prior knowledge, and expertise about the system being con-
trolled (such as the amount of noise, the order of the plant, the number
of unknown parameters, the bandwidth of the parameters’ variation...)
This may be coded as several logic steps or rules, around the adaptive
control algorithm, The resulting composite algorithm is often referred to
as a rule-based control law, with the adaptation scheme being one of the
rules, The design and evaluation of such composite systems is still an
open area of research for nonadaptive as well as adaptive systems,
although adaptive control algorithms form an especially attractive area
of application.
One can conceive of a more complex scenario, in which the plant
to be controlled cannot be easily modeled, either as a linear or nonlinear
system because of the complexity of the physical processes involved. A
controller then has to be built by codifying systematically into rules the
experience gained from operating the system (this is referred to as query
ing and representation of expert knowledge). The rules then serve as @
model of the plant from which the controller is constructed as a rule-
based system, ie. a conjunction of several logic steps and control algo-
rithms. Such a composite de.ign process is called a rule-based expert
controller design. ‘The sophistication and performance of the controller
is dependent on the amount of detail in the model furnished by the
Section 8.2 Future Research 329
pert kaowledge. Adaptation and learning in this framework consist
in reining the rule-based model on the experience salned duseg ie
course of operation ofthe system.
While this framework is extremely attractive ftom a practical poin
control have been implemented, and state ofthe art in learning fos vole
based models is rudimentary. In the context of adaptive contol vec)
interesting study is found in Astrom et ai [1986]. Adapted from theiy
work is Figure 8. illustrating the structure of
F ture of an exper
using an adaptive algorithm. a
rue.
BAseo
EXPERT
sysrem
excrTarion
MONTOR
IDENTIFIER
ALGORITHM
Process
exogenous
iNeuT
conTRoL.
ALconTaM
Figure 8.1: Expert Adaptive Control System
The rule-based system decides, based onthe
ides, based on the lve of excitation, which of
a library of identification algorithms to use and, If nese ee ene
new excitation, It also decides which of a family of eontel ews tre
and communicates its nferencing
its inferencing procedures tothe operator. & taper,
sor provides alarms and interrupts. " Asupervs
Adaptation, Learning, Connectonism and all hose things
While the topics in the title have the sam il
ve the same general philosophical goals,
namely, the understanding, modeling and contol of «given proces we
fields of identification and adaptive control have made the largest strides
in becoming design methodology. by limiting heir vanvens et
course fo a small (ut pracy meaning) cass of sues wih
near or linearizable dynamics, anda finite dimensional sie
‘Learning has, however, been merely parameter updating. a